@codeagora/mcp

CodeAgora MCP server for Claude Code, Cursor, and other MCP-compatible clients.

Install

Release/local smoke uses a packed tarball before publishing:

pnpm --filter @codeagora/mcp build
pnpm --filter @codeagora/mcp pack --dry-run

For published installs, users can run the package with:

npx -y @codeagora/mcp

Client Configuration

{
  "mcpServers": {
    "codeagora": {
      "command": "npx",
      "args": ["-y", "@codeagora/mcp"]
    }
  }
}

For local validation, point your client at the built workspace binary instead:

{
  "mcpServers": {
    "codeagora": {
      "command": "node",
      "args": ["/absolute/path/to/CodeAgora/packages/mcp/dist/index.js"]
    }
  }
}

Environment Variables

Provider keys are required only for live review tools that call LLM providers. Config and stats tools do not require provider keys.

• OPENAI_API_KEY for OpenAI models.
• ANTHROPIC_API_KEY for Anthropic models.
• GROQ_API_KEY for Groq models.
• OPENROUTER_API_KEY for OpenRouter models.
• OPENCODE_API_KEY for OpenCode Go and OpenCode Zen models.
• GITHUB_TOKEN for PR review workflows that fetch or post GitHub data.

Deterministic benchmark CI gates such as pnpm bench:ci do not require live provider keys. Manual live benchmark runs with pnpm bench:fn:run do.

Tools

• review_quick: fast L1-only diff review.
• review_full: full L0-L3 review with debate and head verdict.
• review_pr: PR-oriented review flow. Pass pr_url or pr_number.
• dry_run: estimate diff complexity and review cost from a unified diff.
• explain_session: explain a previous session (repo_path optional).
• get_leaderboard: inspect model ranking data.
• get_stats: inspect review statistics (repo_path optional).
• config_get: read current reviewer configuration (repo_path optional).
• config_set: update supported configuration fields (repo_path optional).

Default MCP review responses are compact to preserve agent context. Request output_format: "json" from review tools when you need the versioned codeagora.review.v1 machine contract.

Common Calls

Use the current workspace root implicitly when the MCP server is already started inside the repo:

{
  "name": "review_quick",
  "arguments": {
    "diff": "diff --git a/src/app.ts b/src/app.ts\n@@ -1,3 +1,4 @@\n+const ready = true;\n"
  }
}

Use staged mode when you do not want to paste a diff:

{
  "name": "review_full",
  "arguments": {
    "staged": true,
    "output_format": "json"
  }
}

Preflight a diff before spending provider tokens:

{
  "name": "dry_run",
  "arguments": {
    "diff": "diff --git a/src/app.ts b/src/app.ts\n@@ -1,3 +1,4 @@\n+const ready = true;\n"
  }
}

Inspect or update config from the same workspace:

{
  "name": "config_get",
  "arguments": {
    "key": "discussion.maxRounds"
  }
}

{
  "name": "config_set",
  "arguments": {
    "key": "discussion.maxRounds",
    "value": 3
  }
}

Explain a prior session without re-running the review:

{
  "name": "explain_session",
  "arguments": {
    "session": "2026-03-19/001"
  }
}

Tool failures use MCP protocol isError: true with a structured JSON body:

{
  "status": "error",
  "code": "INVALID_INPUT",
  "message": "Either diff or staged=true is required"
}

Stable error codes include INVALID_INPUT, INVALID_REPO_PATH, REVIEW_FAILED, REVIEW_PR_FAILED, DRY_RUN_FAILED, CONFIG_GET_FAILED, CONFIG_SET_FAILED, EXPLAIN_SESSION_FAILED, LEADERBOARD_FAILED, and STATS_FAILED.

Troubleshooting

• If the server exits immediately, run pnpm --filter @codeagora/mcp build and confirm packages/mcp/dist/index.js exists.
• If review tools return missing-provider errors, set the provider key for the configured reviewer backend.
• If repo_path is rejected, omit it when you are already inside the workspace; otherwise pass the exact workspace root. Symlinks and paths outside the repo are intentionally rejected.
• The structured error body includes guidance for invalid repo_path, empty diff, and other common retryable failures.
• If a client cannot find the command, use an absolute node path and an absolute dist/index.js path for local beta smoke.